Skip to content

MIRA v0.318.0 — public beta

Pre-release
Pre-release

Choose a tag to compare

View all tags
@tarekedOz tarekedOz released this 25 Jul 18:12
v0.318.0
fdf145b

MIRA v0.318.0 — public beta

MIRA is a self-hostable AI assistant that runs as a single binary. Published by Vexillon under AGPL-3.0-or-later.

Public beta. Supersedes v0.316.1. This release rolls up the changes from v0.317.0 as well (the in-app config panel for apps). No breaking changes — safe to upgrade in place.

Apps framework — sturdier HTTP tools

  • Base URLs work with or without a trailing slash. ${config.base_url}/... template joins now collapse a redundant slash, so an app configured against, say, a Home Assistant instance no longer 404s on double-slash paths.
  • Large responses can't blow the context. An app's HTTP-tool response is now clipped to a small model-facing cap (~24 KB) before it reaches the model.
  • New optional response projection. An app can declaratively select a few dotted fields and a row limit to reduce a large JSON reply to a compact list — no code execution involved.
  • Plus the in-app config panel for apps, folded in from v0.317.0.

Reliability fixes for local models

  • Right-sized context window for local providers. New agent.local_effective_context_tokens (default 8192). When a local provider doesn't declare context_length_tokens, MIRA now budgets to this small, reliable window instead of trusting an advertised 128K that quantized local models don't actually attend to — which keeps prompts in the model's reliable zone and turns on compaction. An explicit context_length_tokens still takes precedence.
  • Degenerate-output guard (on by default). New agent.degeneracy_guard.* settings. A guard wraps the provider chain and aborts pathologically repetitive output early, treats the turn as failed (not persisted, extractors skipped), and counts it as a provider error so fallback can kick in.
  • Guardian triage model falls back sensibly. An empty guardian.triage_model now resolves to the provider's configured default model instead of sending an empty model name.
  • Verified and closed the Just-in-Time Tools find_tools hint (shipped in v0.316.1).

Install

# Linux / macOS
curl -fsSL https://get.vexillon.ai/install.sh | sh
# Windows
irm https://get.vexillon.ai/install.ps1 | iex

Targets

  • x86_64-unknown-linux-gnu
  • aarch64-unknown-linux-gnu
  • x86_64-apple-darwin
  • aarch64-apple-darwin
  • x86_64-pc-windows-msvc

Verify

All artifacts are signed with minisign. Verify the checksum file, then the archive checksums:

minisign -Vm SHA256SUMS -P RWQX8PcjZB3S8i1V9AkZgUKoIy+eZWRzaEHKvBW9U/mkMP9Vp0B8xOaO
sha256sum -c SHA256SUMS

Trusted comment on a valid signature reads: signed by vexillon.ai — MIRA 0.318.0

Assets 14
Loading